Commodity Image Retrieval Based on Convolutional Neural Network and Late Fusion

Abstract

Commodity image retrieval has great research value in the field of e-commerce. However, at the same time, the diversity of commodity images makes the effectiveness of commodity retrieval often unsatisfactory. For image retrieval, the question of whether a given query image is the best query result for the underlying framework using a convolutional neural network as a feature extractor is often neglected. Inspired by the superiority of deep learning in image content understanding and powerful image feature extraction, this paper proposes a new deep learning-based framework for commodity image retrieval to make commodity image retrieval have a better retrieval effect. The framework achieves an efficient retrieval of images by combining deep Convolutional Neural Networks for the late fusion of images at the scoring level. A database of 35 categories of commodities containing 3500 images was created for experimental validation in this paper. Experiments comparing the performance between the frameworks using our dataset show that this paper’s proposed framework has higher retrieval accuracy than the base framework.